Initialize attention backends and forward metadata
(
self,
)
| 335 | self._empty_cache() |
| 336 | |
| 337 | def _initialize_attn_backend( |
| 338 | self, |
| 339 | ) -> None: |
| 340 | """ |
| 341 | Initialize attention backends and forward metadata |
| 342 | """ |
| 343 | assert len(self.attn_backends) == 0 |
| 344 | |
| 345 | num_heads = self.model_config.num_attention_heads // self.parallel_config.tensor_parallel_size |
| 346 | self.model_config.kv_num_heads = max( |
| 347 | 1, |
| 348 | int(self.model_config.num_key_value_heads) // self.parallel_config.tensor_parallel_size, |
| 349 | ) |
| 350 | head_dim = self.model_config.head_dim |
| 351 | |
| 352 | # Initialize AttentionBackend buffers |
| 353 | encoder_block_shape_q = 64 |
| 354 | decoder_block_shape_q = 16 |
| 355 | |
| 356 | self.model_inputs["decoder_batch_ids"] = paddle.zeros_like(self.target_model_inputs["decoder_batch_ids"]) |
| 357 | self.model_inputs["decoder_tile_ids_per_batch"] = paddle.zeros_like( |
| 358 | self.target_model_inputs["decoder_tile_ids_per_batch"] |
| 359 | ) |
| 360 | if current_platform.is_xpu() or current_platform.is_maca(): |
| 361 | self.model_inputs["decoder_num_blocks_cpu"] = paddle.zeros_like( |
| 362 | self.target_model_inputs["decoder_num_blocks_cpu"] |
| 363 | ).cpu() |
| 364 | else: |
| 365 | self.model_inputs["decoder_num_blocks_cpu"] = paddle.zeros_like( |
| 366 | self.target_model_inputs["decoder_num_blocks_cpu"] |
| 367 | ).pin_memory() |
| 368 | self.model_inputs["decoder_num_blocks_device"] = paddle.zeros_like( |
| 369 | self.target_model_inputs["decoder_num_blocks_device"] |
| 370 | ) |
| 371 | self.model_inputs["decoder_chunk_size_device"] = paddle.zeros_like( |
| 372 | self.target_model_inputs["decoder_chunk_size_device"] |
| 373 | ) |
| 374 | self.model_inputs["max_len_tensor_cpu"] = paddle.zeros_like( |
| 375 | self.target_model_inputs["max_len_tensor_cpu"] |
| 376 | ).cpu() |
| 377 | |
| 378 | self.model_inputs["encoder_batch_ids"] = paddle.zeros_like(self.target_model_inputs["encoder_batch_ids"]) |
| 379 | self.model_inputs["encoder_tile_ids_per_batch"] = paddle.zeros_like( |
| 380 | self.target_model_inputs["encoder_tile_ids_per_batch"] |
| 381 | ) |
| 382 | self.model_inputs["encoder_num_blocks_x_cpu"] = paddle.zeros_like( |
| 383 | self.target_model_inputs["encoder_num_blocks_x_cpu"] |
| 384 | ).cpu() |
| 385 | self.model_inputs["kv_batch_ids"] = paddle.zeros_like(self.target_model_inputs["kv_batch_ids"]) |
| 386 | self.model_inputs["kv_tile_ids_per_batch"] = paddle.zeros_like( |
| 387 | self.target_model_inputs["kv_tile_ids_per_batch"] |
| 388 | ) |
| 389 | self.model_inputs["kv_num_blocks_x_cpu"] = paddle.zeros_like( |
| 390 | self.target_model_inputs["kv_num_blocks_x_cpu"] |
| 391 | ).cpu() |
| 392 | |
| 393 | # Get the attention backend |
| 394 | attn_cls = get_attention_backend() |
no test coverage detected